Mao Keli, Yang Yahan, Guo Chong, Zhu Yi, Chen Chuan, Chen Jingchang, Liu Li, Chen Lifei, Mo Zijun, Lin Bingsen, Zhang Xinliang, Li Sijin, Lin Xiaoming, Lin Haotian
State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University, Guangzhou, China.
Department of Molecular and Cellular Pharmacology, University of Miami Miller School of Medicine, Miami, FL, USA.
Ann Transl Med. 2021 Mar;9(5):374. doi: 10.21037/atm-20-5442.
Strabismus affects approximately 0.8-6.8% of the world's population and can lead to abnormal visual function. However, Strabismus screening and measurement are laborious and require professional training. This study aimed to develop an artificial intelligence (AI) platform based on corneal light-reflection photos for the diagnosis of strabismus and to provide preoperative advice.
An AI platform consisting of three deep learning (DL) systems for strabismus diagnosis, angle evaluation, and operation plannings based on corneal light-reflection photos was trained and retrospectively validated using a retrospective development data set obtained between Jan 1, 2014, and Dec 31, 2018. Corneal light-reflection photos were collected to train the DL systems for strabismus screening and deviation evaluations in the horizontal strabismus while concatenated images (each composed of two photos representing different gaze states) were procured to train the DL system for operative advice regarding exotropia. The AI platform was further prospectively validated using a prospective development data set captured between Sep 1, 2019, and Jun 10, 2020.
In total, 5,797 and 571 photos were included in the retrospective and prospectively development data sets, respectively. In the retrospective test sets, the screening system detected strabismus with a sensitivity of 99.1% [95% confidence interval (95% CI), 98.1-99.7%], a specificity of 98.3% (95% CI, 94.6-99.5%), and an AUC of 0.998 (95% CI, 0.993-1.000, P<0.001). Compared to the angle measured by the perimeter arc, the deviation evaluation system achieved a level of accuracy of ±6.6º (95% LoA) with a small bias of 1.0º. Compared to the real design, the operation advice system provided advice regarding the target angle within ±5.5º (95% LoA). Regarding strabismus in the prospective test set, the AUC was 0.980. The platform achieved a level of accuracy of ±7.0º (95% LoA) in the deviation evaluation and ±6.1º (95% LoA) in the target angle suggestion.
The AI platform based on corneal light-reflection photos can provide reliable references for strabismus diagnosis, angle evaluation, and surgical plannings.
斜视影响着全球约0.8%-6.8%的人口，并可能导致视觉功能异常。然而，斜视的筛查和测量工作繁重，且需要专业培训。本研究旨在开发一个基于角膜反光照片的人工智能（AI）平台，用于斜视诊断并提供术前建议。
使用2014年1月1日至2018年12月31日期间获取的回顾性开发数据集，对一个由三个深度学习（DL）系统组成的AI平台进行训练和回顾性验证，这三个系统分别用于基于角膜反光照片的斜视诊断、角度评估和手术规划。收集角膜反光照片以训练DL系统用于水平斜视的筛查和斜视度评估，同时获取拼接图像（每张由代表不同注视状态的两张照片组成）来训练DL系统以提供外斜视手术建议。使用2019年9月1日至2020年6月10日期间获取的前瞻性开发数据集对AI平台进行进一步的前瞻性验证。
回顾性和前瞻性开发数据集中分别纳入了5797张和571张照片。在回顾性测试集中，筛查系统检测斜视的灵敏度为99.1%[95%置信区间（95%CI），98.1%-99.7%]，特异度为98.3%（95%CI，94.6%-99.5%），曲线下面积（AUC）为0.998（95%CI，0.993-1.000，P<0.001）。与弧形视野计测量的角度相比，斜视度评估系统的准确度达到±6.6°（95%一致性界限），偏差较小，为1.0°。与实际设计相比，手术建议系统给出的目标角度建议在±5.5°（95%一致性界限）范围内。在前瞻性测试集中，斜视诊断的AUC为0.980。该平台在斜视度评估中的准确度达到±7.0°（95%一致性界限），在目标角度建议方面达到±6.1°（95%一致性界限）。
基于角膜反光照片的AI平台可为斜视诊断、角度评估和手术规划提供可靠参考。
